ROCm implementation of the Preconditioned Conjugate Gradient method with Incomplete Cholesky (IC0) preconditioner for solving symmetric positive definite linear systems.

• ROCm 6.4+ or ROCm 7.x
• AMD GPU with ROCm support

module load rocm/7.1.1   # or rocm/6.4.0
make

This builds two executables:

• pcg_laplacian - Test driver with generated 3D Laplacian matrix
• pcg_mtx - Driver for Matrix Market format input files

# Default 20x20x20 grid
./pcg_laplacian

# Custom grid size and solver parameters
./pcg_laplacian <nx> <ny> <nz> [tol] [maxiter]

# Example: 50x50x50 grid with tolerance 1e-10
./pcg_laplacian 50 50 50 1e-10 1000

./pcg_mtx <matrix.mtx> [options]

Options:
  -tol <value>      Convergence tolerance (default: 1e-10)
  -maxiter <value>  Maximum iterations (default: 10000)
  -rhs <file.mtx>   Right-hand side vector in Matrix Market format
                    (default: vector of all ones)

Example:

./pcg_mtx thermal2.mtx -tol 1e-8 -maxiter 2000
./pcg_mtx matrix.mtx -rhs rhs.mtx -tol 1e-10

The implementation uses:

• Incomplete Cholesky IC(0) factorization via rocsparse_dcsric0
• Sparse triangular solves via rocsparse_spsv for preconditioning
• Sparse matrix-vector product via rocsparse_spmv
• rocBLAS for vector operations (dot products, scaling, axpy)

• pcg_ic.h - Header file with function declarations
• pcg_ic.cpp - PCG algorithm implementation
• driver.cpp - Laplacian test driver
• driver_mtx.cpp - Matrix Market driver

The code includes version detection macros to handle API differences between ROCm 6.4 and 7.x, suppressing deprecation warnings while maintaining backward compatibility.